An AI-based method for predicting the ecological governance effects of photovoltaic zones

Abstract

This invention discloses an artificial intelligence-based method for predicting the ecological governance effect of photovoltaic areas, which relates to the field of ecological governance technology. The method includes the following steps: dividing the photovoltaic area into a two-dimensional grid and deploying sensors to collect soil data for key areas; obtaining soil data for photovoltaic grid points using an inverse distance weighting method; obtaining red and near-infrared reflectance from UAV remote sensing imagery and calculating vegetation indices; obtaining a comprehensive tensor for photovoltaic grid points by combining the soil data and vegetation indices; inputting the comprehensive tensor into a convolutional neural network model based on a time attention mechanism to output an ecological feature evolution tensor; inputting the ecological feature evolution tensor into an ecological governance effect prediction model to output an ecological governance prediction tensor; and optimizing the parameters of the ecological governance effect prediction model by calculating the difference between the real-time ecological governance tensor and the predicted ecological governance tensor.

Description

Technical Field

[0001] This invention relates to the field of ecological governance technology, specifically to a method for predicting the ecological governance effects of photovoltaic areas based on artificial intelligence. Background Technology

[0002] With the global energy structure shifting towards clean energy, the photovoltaic industry is experiencing rapid development. While the construction of large-scale photovoltaic zones promotes green energy, it also faces challenges in ecological protection and governance. The dynamic changes in ecological elements such as soil quality, vegetation cover, and microclimate in photovoltaic zones directly affect regional ecological stability. Therefore, accurate prediction of the ecological governance effects of photovoltaic zones has become a key link in ensuring the coordinated advancement of sustainable development of the photovoltaic industry and ecological protection.

[0003] The evaluation of the ecological governance effect in photovoltaic areas relies heavily on traditional monitoring and analysis methods. This involves manually deploying sensors to collect data such as soil moisture and temperature, combining this with regular on-site sampling to obtain vegetation growth information, and using statistical models to analyze the soil moisture and temperature data in order to predict the ecological governance effect.

[0004] However, existing methods lack the ability to collaboratively model time series patterns when dealing with the dynamic evolution relationships of multi-dimensional parameters such as soil and vegetation. They are difficult to effectively integrate the dynamic change patterns of ecological data such as soil and vegetation in the time dimension, resulting in low accuracy in predicting the effects of ecological governance and failing to meet the needs of refined governance in large-scale photovoltaic areas. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based method for predicting the ecological governance effects of photovoltaic areas, thereby resolving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence, comprising the following steps:

[0007] Step S1: Divide the photovoltaic area into a grid to obtain a two-dimensional grid of the photovoltaic area; deploy sensors in the two-dimensional grid of the photovoltaic area and collect soil data of key areas through the sensors; based on the soil data of key areas, fill in each point in the two-dimensional grid of the photovoltaic area using the inverse distance weighting method to calculate the soil data of the photovoltaic grid points;

[0008] Step S2: The photovoltaic area is scanned by a drone equipped with a multispectral camera to obtain remote sensing image data. The red band reflectance and near-infrared band reflectance are obtained by frequency band decomposition of the remote sensing image data. The vegetation index of the photovoltaic grid points is calculated by combining the red band reflectance and near-infrared band reflectance. The photovoltaic grid point comprehensive tensor is obtained by combining the soil data and the vegetation index of the photovoltaic grid points.

[0009] Step S3: Construct a convolutional neural network model based on the time attention mechanism, input the photovoltaic grid point comprehensive tensor into the convolutional neural network model, and output the ecological feature evolution tensor;

[0010] Step S4: Construct an ecological governance effect prediction model, input the ecological feature evolution tensor into the ecological governance effect prediction model, and output the ecological governance prediction tensor;

[0011] Step S5: Collect real-time ecological governance data of the photovoltaic area to obtain the real-time ecological governance tensor; calculate the mean square error of ecological governance by calculating the parameter difference between the real-time ecological governance tensor and the predicted ecological governance tensor; calculate the channel similarity index of ecological governance by calculating the channel similarity between the real-time ecological governance tensor and the predicted ecological governance tensor.

[0012] Step S6: By combining the mean square error of ecological governance and the similarity index of ecological governance channels, the ecological governance error index is calculated; based on the ecological governance error index, the parameters of the ecological governance effect prediction model are optimized, and finally the optimized ecological governance effect prediction model is obtained, thereby realizing the prediction of ecological governance effect.

[0013] Preferably, the step of filling in each point in the two-dimensional grid of the photovoltaic area using the inverse distance weighting method based on the soil data of the key area to calculate the soil data of the photovoltaic grid points includes the following specific steps:

[0014] The soil data for each point in the two-dimensional grid of the photovoltaic area is calculated by filling in the grid using the inverse distance weighting method, based on the j-th feature value in the soil data of the key area:

[0015]

[0016] in, This represents the j-th feature value of the soil data at the photovoltaic grid point (h, w) in the two-dimensional grid of the photovoltaic area. Let (h, w) be the distance from the coordinates (h, w) to the soil data of the b-th key region, where b represents the index of the soil data of the b-th key region, and B represents the total number of soil data in the key regions. This represents the j-th feature value of the soil data in the b-th key area.

[0017] Preferably, the step of obtaining red band reflectance and near-infrared band reflectance by frequency band decomposition of the remote sensing image data includes the following steps:

[0018] In UAV remote sensing image data, each pixel records the DN value for each band. The DN value is then converted into surface reflectance.

[0019]

[0020] in, For surface reflectance, The radiance is converted from the DN value using sensor gain and offset. The distance between the Earth and the Sun. Solar irradiance, The solar zenith angle;

[0021] Based on surface reflectance data, the reflectance in the near-infrared band (700-1100nm) of UAV remote sensing imagery is extracted. Reflectivity in the red light band of the 600-700nm range .

[0022] Preferably, the step of calculating the vegetation index of the photovoltaic grid points by combining the red light band reflectance and the near-infrared band reflectance includes the following steps:

[0023] By combining the red light band reflectance and the near-infrared band reflectance, the vegetation index of the photovoltaic grid points is calculated:

[0024]

[0025] in, This represents the vegetation index at the photovoltaic grid point with coordinates (h,w). This represents the near-infrared reflectance at coordinates (h, w). This represents the reflectivity in the red light band at coordinates (h, w).

[0026] Preferably, the construction of the convolutional neural network model based on the temporal attention mechanism includes the following steps:

[0027] A convolutional neural network model based on a temporal attention mechanism is constructed. The convolutional neural network model includes: a spatial feature extraction branch, a temporal feature extraction branch, and a feature fusion and compression module.

[0028] The spatial feature extraction branch includes: convolutional layer 1, convolutional layer 2, spatial attention module, max pooling layer, convolutional layer 3, and convolutional layer 4;

[0029] Convolutional layer 1: 2D convolutional layer, 3 convolutional kernels 3. Output channels: 64, stride: 1, padding: 1, ReLU activation function;

[0030] Convolutional layer 2: 2D convolutional layer, convolutional kernel 3 3. Output channels: 64, stride: 1, padding: 1, ReLU activation function;

[0031] Spatial Attention Module: Channel Attention: Global Average Pooling - Global Max Pooling - Shared Multilayer Perceptron (MLP) - Sigmoid activation to generate channel weights; Spatial Attention: Using channel average and max pooling to obtain 2-channel features - 7×7 convolution - Sigmoid activation to generate spatial weights;

[0032] Max pooling layer: pooling window 2 2, step size is 2;

[0033] Convolutional layer 3: 2D convolutional layer, 3 convolutional kernels 3. Output channels: 128, stride: 1, padding: 1, ReLU activation function;

[0034] Convolutional layer 4: 2D convolutional layer, with 3 convolutional kernels. 3. Output channels: 128, stride: 1, padding: 1, ReLU activation function;

[0035] Spatial Attention Module: Channel Attention: Global Average Pooling - Global Max Pooling - Shared Multilayer Perceptron (MLP) - Sigmoid activation to generate channel weights; Spatial Attention: Using channel average and max pooling to obtain 2-channel features - 7×7 convolution - Sigmoid activation to generate spatial weights;

[0036] Max pooling layer: pooling window 2 2, step size is 2;

[0037] The temporal feature extraction branch consists of: a temporal attention layer, a fully connected layer, a feedforward neural network, a global average pooling layer, and a max pooling layer. The temporal attention layer employs a multi-head attention mechanism, with each head performing Q, K, and V transformations. The fully connected layer performs linear transformations. The max pooling layer is a 1D max pooling layer with a pooling window of 1*4 and a stride of 4.

[0038] Feature fusion and compression module: temporal average pooling layer, splicing layer, convolutional layer 5; convolutional layer 5: 1×1 convolutional kernel, 128 output channels.

[0039] Preferably, the step of inputting the photovoltaic grid point integrated tensor into the convolutional neural network model and outputting the ecological feature evolution tensor includes the following specific steps:

[0040] The photovoltaic grid point synthesis tensor X is input into the convolutional neural network model, and the ecological feature evolution tensor is output. , Where H' and W' are the number of grid rows and columns output by the convolutional neural network model, T is the time step, and D is the ecological parameter channel output by the convolutional neural network model.

[0041] Preferably, the construction of the ecological governance effect prediction model includes the following specific steps:

[0042] An ecological governance effect prediction model is constructed, which includes a 3D dilated convolution module and a hierarchical temporal module, a multi-scale feature fusion layer, a 3D high-precision upsampling layer, and a prediction output layer.

[0043] The 3D dilated convolution module includes: dilated convolution layer 1, dilated convolution layer 2, and a spatial attention module;

[0044] Hollow convolution layer 1: Convolution kernel 3 3 3. Drillion ratio (1,2,2), number of output channels 128; dilated convolutional layer 2: convolutional kernel 3 3 2. Hollow rate (1,2,2), number of output channels 128; Spatial attention module: Channel attention: global average pooling, global max pooling, shared multilayer perceptron (MLP), sigmoid activation to generate channel weights; Spatial attention: using channel average and max pooling to obtain 2-channel features, 7×7 convolution, sigmoid activation to generate spatial weights.

[0045] The hierarchical temporal module includes: a short-term branch and a long-term branch. The short-term branch adopts a 3-layer convolutional long short-term memory network 3D-ConvLSTM, and the long-term branch adopts a 4-head attention layer and a feedforward network.

[0046] 3D high-precision upsampling layer: transposed convolution: 3×3 kernels, stride 2, 128 channels, output 100×100×128; double transposed convolution: 3×3 kernels, stride 2, 64 channels;

[0047] Prediction Output Layer: Ecological Governance Prediction Tensor , H represents the number of grid rows, W represents the number of grid columns, and K represents the ecological parameter channels. To predict the time step.

[0048] Preferably, the step of calculating the mean square error of ecological governance by calculating the parameter difference between the real-time ecological governance tensor and the predicted ecological governance tensor includes the following specific steps:

[0049] The mean square error of ecological governance is calculated by determining the parameter difference between the real-time ecological governance tensor and the predicted ecological governance tensor.

[0050]

[0051] RMSE represents the mean square error of ecological governance. To predict the time step, t is the time step index, H is the number of grid rows, W is the number of grid columns, h is the index of the grid row number, and w is the index of the grid column number. This represents the ecological governance prediction tensor with h rows and w columns. This represents a real-time ecological governance tensor with h rows and w columns.

[0052] Preferably, the step of calculating the ecological governance channel similarity index by calculating the channel difference between the real-time ecological governance tensor and the predicted ecological governance tensor includes the following specific steps:

[0053] By calculating the channel similarity between the real-time ecological governance tensor and the predicted ecological governance tensor, the ecological governance channel similarity index is obtained:

[0054]

[0055] Among them, SSIM is the similarity index of ecological governance channels. To predict the time step, t is the time step index, K is the number of channels, and k is the channel index. Let be the mean of the predicted values ​​for the k-th channel at time step t. Let be the average of the measured values ​​of the k-th channel at time step t. Let Variance be the predicted value variance of the k-th channel at time step t. Let V be the variance of the measured values ​​of the k-th channel at time step t. Let be the covariance of the k-th channel at time step t. To prevent the mean from being divided by zero, the default value is... , The variance is divided by zero constant, which is the default value. .

[0056] Preferably, the step of calculating the ecological governance error index by combining the mean square error of ecological governance and the similarity index of ecological governance channels includes the following specific steps:

[0057] By combining the mean square error of ecological governance and the similarity index of ecological governance channels, the ecological governance error index is calculated:

[0058]

[0059] Among them, EGDV is the ecological governance error index. For normalized root mean square error, = MAXSE represents the maximum root mean square error, H is the number of grid rows, W is the number of grid columns, h is the index of the grid row, and w is the index of the grid column. The mean square error weighting coefficient is... The channel similarity index weighting coefficient, + =1.

[0060] This invention provides an artificial intelligence-based method for predicting the ecological governance effects of photovoltaic areas, involving machine learning and deep learning technologies, which has the following beneficial effects:

[0061] (1) By combining the soil data and vegetation index of the photovoltaic grid points, a photovoltaic grid point comprehensive tensor is formed, which integrates the soil attributes and vegetation status of each grid point into a unified data structure without retaining the local features of each grid point.

[0062] (2) By combining the root mean square error of ecological governance and the similarity index of ecological governance channels, the calculated ecological governance error index provides a multi-dimensional model performance evaluation index. This ecological governance error index not only focuses on the absolute accuracy of the predicted values, but also emphasizes the evaluation of the performance of the predicted results in terms of spatial structure consistency and the matching degree of multi-factor coordinated change trends. It can more realistically reflect the comprehensive performance of the model when simulating the dynamics of complex ecosystems, avoid the spatial prediction distortion problem that may be masked by a single RMSE index, and provide a more reliable basis for the accurate optimization of the model.

[0063] (3) Based on the calculated ecological governance error index, the parameters of the ecological governance effect prediction model are optimized, and the optimized ecological governance effect prediction model is finally obtained. The model prediction ability is continuously self-evolved and adapted to environmental dynamics. The ecological governance error index is used as a feedback signal to drive the automatic adjustment of model parameters such as neural network weights, learning rate, regularization coefficient, etc. It can correct the systematic deviations or deficiencies found in the model prediction in real time based on the latest monitoring data feedback. Attached Figure Description

[0064] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1This is a flowchart illustrating the steps of a method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence, as proposed in this invention.

[0066] Figure 2 This is a step hierarchy diagram for obtaining the photovoltaic grid point tensor in an artificial intelligence-based method for predicting the ecological governance effect of photovoltaic areas proposed in this invention;

[0067] Figure 3 This is a step hierarchy diagram of the optimized ecological governance effect prediction model obtained in the artificial intelligence-based photovoltaic area ecological governance effect prediction method proposed in this invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Please see Figures 1-3 This invention provides a technical solution: a method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence.

[0070] Step S1 involves dividing the photovoltaic area into a grid to obtain a two-dimensional grid for the photovoltaic area; sensors are deployed in the two-dimensional grid of the photovoltaic area to collect soil data for key areas; based on the soil data for key areas, each point in the two-dimensional grid of the photovoltaic area is filled in using the inverse distance weighting method to calculate the soil data for the photovoltaic grid points.

[0071] Within the selected photovoltaic area, the area is divided into a grid, resulting in a two-dimensional grid. The grid resolution can be determined by the terrain; the default grid resolution is 10m for flat areas, while for slopes or drainage ditches, finer monitoring can be performed with a grid resolution of 1-5m. Sensors are deployed in key areas of the two-dimensional grid, including bare areas below the grid, surrounding vegetation restoration areas, and areas prone to soil erosion. Soil moisture sensors, surface temperature sensors, and wind speed sensors are deployed in these key areas. The soil moisture sensors are buried at different depths in the soil to collect soil moisture content data at different depths. The surface temperature sensors are fixed near the surface to record changes in surface temperature in real time. The wind speed sensors are installed at a height of approximately 1.5 meters above the ground to continuously monitor wind speed within the area, ultimately obtaining soil data for the key areas.

[0072] The soil data for each point in the two-dimensional grid of the photovoltaic area is calculated by filling in the grid using the inverse distance weighting method, based on the j-th feature value in the soil data of the key area:

[0073]

[0074] in, This represents the j-th feature value of the soil data at the photovoltaic grid point (h, w) in the two-dimensional grid of the photovoltaic area. Let (h, w) be the distance from the coordinates (h, w) to the soil data of the b-th key region, where b represents the index of the soil data of the b-th key region, and B represents the total number of soil data in the key regions. This represents the j-th feature value of the soil data in the b-th key area.

[0075] It should be noted that the soil data of each grid point in the two-dimensional grid of the entire photovoltaic area is calculated based on the key area soil data collected by the sensor and the inverse distance weighting method. The grid points closer to the sensor location are given higher weights (inverse square distance). Data is filled in for areas where no sensors are deployed, which solves the practical limitation that the sensors cannot fully cover the photovoltaic area and ensures the continuity and smooth transition of spatial data.

[0076] The soil data of each grid point in the photovoltaic grid is finally obtained by using the inverse distance weighting method.

[0077] Step S2: Scan the photovoltaic area using a drone equipped with a multispectral camera to obtain remote sensing image data. Decompose the remote sensing image data into frequency bands to obtain the red band reflectance and near-infrared band reflectance. Calculate the vegetation index of the photovoltaic grid points by combining the red band reflectance and near-infrared band reflectance. Obtain the photovoltaic grid point comprehensive tensor by combining the soil data and vegetation index of the photovoltaic grid points.

[0078] Before scanning the photovoltaic area, the flight range needs to be delineated based on the regional topography, photovoltaic panel distribution, and key ecological monitoring areas. The flight altitude is typically set at 50-100 meters to balance image resolution and coverage efficiency. A period of stable lighting, such as 10:00 AM to 2:00 PM, should be selected to avoid spectral interference caused by direct sunlight or cloudy days. Subsequently, the drone's battery life, GPS positioning accuracy, and multispectral camera status are checked. Radiometric calibration of the camera is performed to ensure accurate reflectivity data for different wavelengths, and geometric calibration is performed to reduce the impact of lens distortion. During flight, the drone automatically cruises along a preset route, and the multispectral camera simultaneously collects images covering vegetation-sensitive wavelengths, including visible light and near-infrared bands, to facilitate subsequent identification of vegetation growth status. The flight trajectory and data storage are monitored in real time, with a focus on scanning key areas such as bare ground between photovoltaic panels and vegetation restoration areas. After the flight, the equipment is retrieved and the remote sensing image data is exported.

[0079] It should be noted that if the pixel resolution of the UAV image is inconsistent with the resolution of the photovoltaic grid, bilinear interpolation (pixel < grid) or pixel aggregation (pixel > grid) methods are used to ensure that each subsequent photovoltaic grid point corresponds to one vegetation index value.

[0080] In UAV remote sensing image data, each pixel records the DN value for each band. The DN value is then converted into surface reflectance to eliminate sensor noise and atmospheric effects.

[0081]

[0082] in, For surface reflectance, The radiance is converted from the DN value using sensor gain and offset. The distance between the Earth and the Sun. Solar irradiance, The solar zenith angle is calculated using the imaging time and geographical location.

[0083] It should be noted that the radiance L is calculated as L = G * DN + B, where DN is the original numerical value, G is the gain coefficient obtained through laboratory calibration, and B is the offset.

[0084] For the radiometrically corrected surface reflectance data, the reflectance values ​​in the pre-defined near-infrared band (typically corresponding to the 700-1100nm range) and red band (typically corresponding to the 600-700nm range) of the UAV remote sensing imagery are extracted and denoted as follows: and , This represents the near-infrared reflectance at coordinates (h, w). This represents the reflectivity in the red light band at coordinates (h, w).

[0085] By combining the red light band reflectance and the near-infrared band reflectance, the vegetation index of the photovoltaic grid points is calculated:

[0086]

[0087] in, This represents the vegetation index at the photovoltaic grid point with coordinates (h,w). This represents the near-infrared reflectance at coordinates (h, w). This represents the reflectivity in the red light band at coordinates (h, w).

[0088] It should be noted that the vegetation index of the photovoltaic grid points is calculated by combining the red light band reflectance and the near-infrared band reflectance. The red light band (around 600nm-700nm) reflects chlorophyll absorption characteristics, while the near-infrared band (700-1100nm) is highly sensitive to leaf cell structure. The standardized ratio of their reflectances... It can effectively eliminate interference from light variations and soil background, highlighting vegetation physiological characteristics, thereby converting complex spectral information into vegetation indices. The value range is [-1, 1], and the larger the positive value, the higher the vegetation coverage.

[0089] After standardizing the soil data and vegetation index of the photovoltaic grid points according to their maximum and minimum values, the data are combined to form the photovoltaic grid point comprehensive tensor O, O H represents the number of grid rows, W represents the number of grid columns, K represents the ecological parameter channel, and T represents the time step. The ecological parameter channel includes soil data and vegetation index of photovoltaic grid points.

[0090] Step S3: Construct a convolutional neural network model based on the time attention mechanism, input the photovoltaic grid point comprehensive tensor into the convolutional neural network model, and output the ecological feature evolution tensor.

[0091] A convolutional neural network model based on a temporal attention mechanism is constructed. The convolutional neural network model includes a spatial feature extraction branch, a temporal feature extraction branch, and a feature fusion and compression module.

[0092] The spatial feature extraction branch includes: convolutional layer 1, convolutional layer 2, spatial attention module, max pooling layer, convolutional layer 3, and convolutional layer 4.

[0093] Convolutional layer 1: 2D convolutional layer, 3 convolutional kernels 3. Output channels: 64, step size: 1, padding: 1, ReLU activation function.

[0094] Convolutional layer 2: 2D convolutional layer, convolutional kernel 3 3. Output channels: 64, step size: 1, padding: 1, ReLU activation function.

[0095] Spatial attention module: Channel attention: global average pooling - global max pooling - shared multilayer perceptron (MLP) (2 layers, with 64 and 32 neurons) - Sigmoid activation to generate channel weights; Spatial attention: using channel average and max pooling to obtain 2-channel features - 7×7 convolution - Sigmoid activation to generate spatial weights.

[0096] Max pooling layer: pooling window 2 2, step size is 2.

[0097] Convolutional layer 3: 2D convolutional layer, 3 convolutional kernels 3. Output channels: 128, step size: 1, padding: 1, ReLU activation function.

[0098] Convolutional layer 4: 2D convolutional layer, with 3 convolutional kernels. 3. Output channels: 128, step size: 1, padding: 1, ReLU activation function.

[0099] Spatial attention module: Channel attention: global average pooling - global max pooling - shared multilayer perceptron (MLP) (2 layers, with 64 and 32 neurons) - Sigmoid activation to generate channel weights; Spatial attention: using channel average and max pooling to obtain 2-channel features - 7×7 convolution - Sigmoid activation to generate spatial weights.

[0100] Max pooling layer: pooling window 2 2, step size 2.

[0101] Spatial global average pooling is performed on the integrated tensor of the photovoltaic grid points, and flattened into K. The time sequence T is input to the time feature extraction branch. The time feature extraction branch consists of: a time attention layer, a fully connected layer, a feedforward neural network, a global average pooling layer, and a max pooling layer. The time attention layer (with 4 heads, and a total dimension of Q and K) is... With a dropout rate of 0.1%, a multi-head attention mechanism is employed, with each head performing Q, K, and V transformations; fully connected layers perform linear transformations; max pooling layers: 1D max pooling layers with a pooling window of 1*4, a stride of 4, and a scope encompassing the T-dimensional time series, compressing the time steps to... .

[0102] Feature fusion and compression module: Temporal feature extraction branch output 1 1 128, extended via broadcast operation to The output is concatenated with the spatial feature extraction branch, and then compressed through convolutional layer 5. The sequence includes a temporal average pooling layer, a concatenation layer, and convolutional layer 5. Convolutional layer 5 has a 1×1 kernel and 128 output channels.

[0103] The photovoltaic grid point synthesis tensor X is input into the convolutional neural network model, and the ecological feature evolution tensor is output. , Where H' and W' are the number of grid rows and columns in the output of the convolutional neural network model, H' = W'= (through two convolutional neural network models) (2 pooling), T is the time step, and D is the ecological parameter channel (128 channels) output by the convolutional neural network model.

[0104] It should be noted that a convolutional neural network model based on a temporal attention mechanism (including spatial feature extraction and temporal feature extraction branches) is constructed to perform deep spatiotemporal feature fusion on the input photovoltaic grid point comprehensive tensor (containing soil data and vegetation indices for each grid point at time step T). Its output is an ecological feature evolution tensor. Each grid point (h', w') at each time step t∈T generates a 128-dimensional abstract feature vector (such as vegetation response pattern or water diffusion trend), providing a high-dimensional dynamic representation for subsequent predictions.

[0105] Step S4: Construct an ecological governance effect prediction model, input the ecological feature evolution tensor into the ecological governance effect prediction model, and output the ecological governance prediction tensor.

[0106] An ecological governance effect prediction model is constructed, which includes a 3D dilated convolution module and a hierarchical temporal module, a multi-scale feature fusion layer, a 3D high-precision upsampling layer, and a prediction output layer.

[0107] The 3D dilated convolution module includes: dilated convolution layer 1, dilated convolution layer 2, and a spatial attention module.

[0108] Hollow convolution layer 1: Convolution kernel 3 3 3. Drillion ratio (1,2,2), number of output channels 128; dilated convolutional layer 2: convolutional kernel 3 3 3. Hollow rate (1,2,2), number of output channels 128; Spatial attention module: Channel attention: global average pooling, global max pooling, shared multilayer perceptron (MLP), sigmoid activation to generate channel weights; Spatial attention: use channel average and max pooling to obtain 2-channel features, 7×7 convolution, sigmoid activation to generate spatial weights.

[0109] The evolution tensor of the ecological characteristics After being input into the spatial attention module, the output is a spatial output tensor.

[0110] The hierarchical temporal module includes a short-term branch and a long-term branch. The short-term branch employs a 3-layer convolutional long short-term memory network (3D-ConvLSTM) to output a short-term output tensor.

[0111] Long-term branch: 4-head attention layer and feedforward network, outputting long-term output tensor.

[0112] Multi-scale feature fusion layer:

[0113]

[0114] in, To merge the tensors, S is the spatial output tensor. For spatial feature weights, This indicates a channel splicing operation. For short-term output tensors, For long-term output tensors.

[0115] 3D high-precision upsampling layer: transposed convolution: 3×3 3 cores, stride (1,2,2), 128 channels, output 100×100×128; quadratic transpose convolution: 3×3 3 cores, step size 2, 64 channels.

[0116] Prediction Output Layer: Ecological Governance Prediction Tensor , H represents the number of grid rows, W represents the number of grid columns, and K represents the ecological parameter channels. To predict the time step.

[0117] It should be noted that by constructing a multi-level ecological governance effect prediction model (including a 3D dilated convolution module to capture spatial multi-scale information, and a hierarchical temporal module combining short-term 3D-ConvLSTM and long-term multi-head attention mechanism to model complex temporal dependencies), an ecological governance prediction tensor is output, thereby accurately predicting the evolution trend of key ecological parameters such as vegetation cover and soil moisture in the photovoltaic area in future time steps, providing a dynamic and quantitative scientific basis for ecological governance decisions.

[0118] Step S5: Collect real-time ecological governance data of the photovoltaic area to obtain the real-time ecological governance tensor; calculate the mean square error of ecological governance by calculating the parameter difference between the real-time ecological governance tensor and the predicted ecological governance tensor; calculate the channel similarity index of ecological governance by calculating the channel similarity between the real-time ecological governance tensor and the predicted ecological governance tensor.

[0119] Real-time collection of ecological governance data from photovoltaic areas yields a real-time ecological governance tensor. , H represents the number of grid rows, W represents the number of grid columns, and K represents the ecological parameter channels. To predict the time step.

[0120] It should be noted that the real-time tensor for ecological governance is the same as the prediction tensor for ecological governance. The actual data for each grid point in the H×W grid at each time step includes: soil moisture content data after standardization of maximum and minimum values, surface temperature, wind speed, and vegetation index.

[0121] The mean square error of ecological governance is calculated by determining the parameter difference between the real-time ecological governance tensor and the predicted ecological governance tensor.

[0122]

[0123] RMSE represents the mean square error of ecological governance. To predict the time step, t is the time step index, H is the number of grid rows, W is the number of grid columns, h is the index of the grid row number, and w is the index of the grid column number. This represents the ecological governance prediction tensor with h rows and w columns. This represents a real-time ecological governance tensor with h rows and w columns.

[0124] By calculating the channel similarity between the real-time ecological governance tensor and the predicted ecological governance tensor, the ecological governance channel similarity index is obtained:

[0125]

[0126] Among them, SSIM is the similarity index of ecological governance channels. To predict the time step, t is the time step index, K is the number of channels, and k is the channel index. Let be the mean of the predicted values ​​for the k-th channel at time step t. Let be the average of the measured values ​​of the k-th channel at time step t. Let Variance be the predicted value variance of the k-th channel at time step t. Let V be the variance of the measured values ​​of the k-th channel at time step t. Let be the covariance of the k-th channel at time step t. To prevent the mean from being divided by zero, the default value is... , The variance is divided by zero constant, which is the default value. .

[0127] It should be noted that the mean removal zero constant and the variance removal zero constant are used to avoid the denominator approaching zero due to the overly small means and variances of the predicted values or measured values when calculating the ecological governance channel similarity index SSIM, thus ensuring the stability and rationality of the calculation. Their values are usually related to the dynamic range of ecological parameters and can be set as small constants (such as the square of the parameter maximum value multiplied by a very small coefficient).

[0128] Step S6: By combining the ecological governance mean square error and the ecological governance channel similarity index, calculate the ecological governance error index; optimize the parameters of the ecological governance effect prediction model based on the ecological governance error index, and finally obtain the optimized ecological governance effect prediction model, thereby achieving the prediction of the ecological governance effect.

[0129] By combining the ecological governance mean square error and the ecological governance channel similarity index, calculate the ecological governance error index as follows:

[0130]

[0131] where EGDV is the ecological governance error index, is the normalized root mean square error, = , MAXSE is the maximum value of the root mean square error, H is the number of grid rows, W is the number of grid columns, h is the index of the grid row, w is the index of the grid column, is the mean square error weight coefficient, is the channel similarity index weight coefficient, + = 1.

[0132] It should be noted that the maximum value of the root mean square error is: t (t prediction time step ) corresponding to the maximum value of the ecological governance mean square error.

[0133] It should be noted that the mean square error weight coefficient and the channel similarity index weight coefficient , if quantitative control is required for ecological governance and numerical accuracy needs to be prioritized, then by default is 0.7, is 0.3. It can be automatically adjusted once in each cycle during the training process. For example, when the SSIM improvement rate < RMSE decrease rate, the β weight needs to be increased to strengthen the spatial structure learning; when the SSIM improvement rate ≥ RMSE decrease rate, the α weight needs to be increased to optimize the physical quantity accuracy.

[0134] It should be noted that the Ecological Governance Error Index (EGDV) is calculated by combining the Root Mean Square Error (RMSE) and the Similarity Index for Ecological Governance Channels (SSIM). RMSE reflects the overall numerical deviation between the predicted and real-time values, representing the absolute magnitude of the error; SSIM measures the similarity of various ecological parameter channels, reflecting the matching degree at the feature level. Combining these two metrics comprehensively quantifies the error of the prediction model, avoiding the limitation of single numerical indicators ignoring feature similarity, and compensating for the inability of single similarity indicators to reflect absolute deviation. This provides a more comprehensive basis for parameter optimization of the ecological governance effect prediction model, ensuring that the optimized model can more accurately predict the ecological governance effect of photovoltaic areas.

[0135] Based on the ecological governance error index, the parameters of the ecological governance effect prediction model are optimized to obtain the optimized ecological governance effect prediction model, thereby realizing the prediction of ecological governance effect.

[0136] It should be noted that when optimizing the parameters of the ecological governance effect prediction model based on the Ecological Governance Error Index (EGDV), the goal is to minimize EGDV. The backpropagation algorithm is employed, using the error index as the core indicator of the loss function. The gradients of model parameters such as the kernel weights of dilated convolutional layers, the hidden layer parameters of the ConvLSTM in the hierarchical temporal module, the weight matrix of the Transformer attention head, and the transposed convolutional parameters of the upsampling layer with respect to EGDV are calculated, and the parameters are iteratively updated according to a preset learning rate. Simultaneously, an iteration termination condition is set (e.g., EGDV changes less than a threshold for multiple consecutive iterations or reaching the maximum number of iterations). The changes in EGDV during the optimization process are monitored in real-time using a validation set to avoid overfitting. Ultimately, the model parameters are adjusted to achieve a stable output with minimal difference from real-time data, resulting in the optimized ecological governance effect prediction model.

[0137] It should be noted that, based on the ecological governance error index EGDV, the parameters of the convolutional neural network model in step S3 are also optimized. Through an end-to-end backpropagation mechanism, the gradient of the EGDV loss function is passed layer by layer from the output layer of the ecological governance effect prediction model in step S4 to the convolutional neural network model in step S3: First, the gradient of EGDV with respect to the ecological feature evolution tensor F is calculated. Then, all trainable parameters in the convolutional neural network model are updated based on the chain rule, including the kernel weights and bias terms of the spatial convolutional layers (Conv1-4), as well as the query (Q), key (K), and value (V) projection matrices and multi-head attention weights in the temporal attention module. During the optimization process, an adaptive learning rate strategy (such as the Adam optimizer) is adopted to dynamically adjust the parameter update step size. Gradient clipping (with a threshold set to 1.0) is added for the temporal attention layer to prevent gradient explosion during the temporal feature learning process. Finally, the convolutional neural network model adaptively optimizes its spatiotemporal feature extraction capability, significantly improving the accuracy of capturing dynamic patterns such as vegetation index mutations and seasonal fluctuations in soil moisture.

[0138] This paper proposes an intelligent prediction method for the ecological governance effects of photovoltaic areas based on artificial intelligence, aiming to address the problems of existing technologies relying on a single data source, static models, lacking consideration of the coupled effects of multiple factors such as climate, soil, and vegetation, and insufficient adaptability. This method integrates multi-source heterogeneous data by constructing a multi-dimensional environmental information fusion framework, and designs an automatic extraction mechanism based on key spatiotemporal features, a multi-level spatiotemporal deep learning prediction model, and a model adaptive online update and feedback mechanism. This achieves high-precision and dynamic prediction of the ecological governance effects of photovoltaic areas, such as future vegetation coverage, soil moisture retention, and surface temperature changes, significantly improving the scientific rigor, timeliness, and adaptability to complex environmental changes.

[0139] By combining soil data and vegetation indices from photovoltaic grid points, a comprehensive tensor for photovoltaic grid points is formed. This integrates the soil attributes and vegetation status of each grid point into a unified data structure, without retaining the local features of each grid point.

[0140] By combining the root mean square error (RMSE) of ecological governance and the similarity index of ecological governance channels, an ecological governance error index is calculated, providing a multi-dimensional model performance evaluation metric. This index not only focuses on the absolute accuracy of predicted values ​​but also emphasizes the performance of predicted results in terms of spatial structural consistency and the matching degree of multi-factor collaborative change trends. It more realistically reflects the comprehensive performance of the model when simulating the dynamics of complex ecosystems, avoiding the spatial prediction distortion problems that may be masked by a single RMSE index (e.g., the average predicted value is correct but the spatial distribution is incorrect), thus providing a more reliable basis for precise model tuning.

[0141] Based on the calculated ecological governance error index, the parameters of the ecological governance effect prediction model are optimized, resulting in an optimized model that achieves continuous self-evolution of predictive capabilities and adaptability to environmental dynamics. This process uses the error index as a feedback signal to drive the automatic adjustment of model parameters (such as neural network weights, learning rate, and regularization coefficients). It can instantly correct systematic biases or deficiencies found in model predictions based on the latest monitoring data; through continuous optimization, it enhances the model's generalization ability and long-term predictive stability in complex and non-stationary environments; and it reduces the need for frequent manual model adjustments, enabling intelligent and autonomous operation and performance maintenance of the prediction system with little or no human intervention.

[0142] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0143] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence, characterized in that: Includes the following steps: Step S1: Divide the photovoltaic area into a grid to obtain a two-dimensional grid of the photovoltaic area; deploy sensors in the two-dimensional grid of the photovoltaic area and collect soil data of key areas through the sensors; based on the soil data of key areas, fill in each point in the two-dimensional grid of the photovoltaic area using the inverse distance weighting method to calculate the soil data of the photovoltaic grid points; Step S2: The photovoltaic area is scanned by a drone equipped with a multispectral camera to obtain remote sensing image data. The red band reflectance and near-infrared band reflectance are obtained by frequency band decomposition of the remote sensing image data. The vegetation index of the photovoltaic grid points is calculated by combining the red band reflectance and near-infrared band reflectance. The photovoltaic grid point comprehensive tensor is obtained by combining the soil data and the vegetation index of the photovoltaic grid points. Step S3: Construct a convolutional neural network model based on a time attention mechanism. Input the photovoltaic grid point comprehensive tensor into the convolutional neural network model and output the ecological feature evolution tensor. The construction of the convolutional neural network model based on a time attention mechanism includes the following steps: A convolutional neural network model based on a temporal attention mechanism is constructed. The convolutional neural network model includes: a spatial feature extraction branch, a temporal feature extraction branch, and a feature fusion and compression module. The spatial feature extraction branch includes: convolutional layer 1, convolutional layer 2, spatial attention module, max pooling layer, convolutional layer 3, and convolutional layer 4; Convolutional layer 1: 2D convolutional layer, 3 convolutional kernels 3. Output channels: 64, stride: 1, padding: 1, ReLU activation function; Convolutional layer 2: 2D convolutional layer, convolutional kernel 3 3. Output channels: 64, stride: 1, padding: 1, ReLU activation function; Spatial Attention Module: Channel Attention: Global Average Pooling - Global Max Pooling - Shared Multilayer Perceptron (MLP) - Sigmoid activation to generate channel weights; Spatial Attention: Using channel average and max pooling to obtain 2-channel features - 7×7 convolution - Sigmoid activation to generate spatial weights; Max pooling layer: pooling window 2 2, step size is 2; Convolutional layer 3: 2D convolutional layer, 3 convolutional kernels 3. Output channels: 128, stride: 1, padding: 1, ReLU activation function; Convolutional layer 4: 2D convolutional layer, with 3 convolutional kernels.

3. Output channels: 128, stride: 1, padding: 1, ReLU activation function; Spatial Attention Module: Channel Attention: Global Average Pooling - Global Max Pooling - Shared Multilayer Perceptron (MLP) - Sigmoid activation to generate channel weights; Spatial Attention: Using channel average and max pooling to obtain 2-channel features - 7×7 convolution - Sigmoid activation to generate spatial weights; Max pooling layer: pooling window 2 2, step size is 2; The temporal feature extraction branch consists of: a temporal attention layer, a fully connected layer, a feedforward neural network, a global average pooling layer, and a max pooling layer. The temporal attention layer employs a multi-head attention mechanism, with each head performing Q, K, and V transformations. The fully connected layer performs linear transformations. The max pooling layer is a 1D max pooling layer with a pooling window of 1*4 and a stride of 4. Feature fusion and compression module: temporal average pooling layer, splicing layer, convolutional layer 5; convolutional layer 5: 1×1 convolutional kernel, 128 output channels; Step S4: Construct an ecological governance effect prediction model. Input the ecological feature evolution tensor into the ecological governance effect prediction model and output the ecological governance prediction tensor. The construction of the ecological governance effect prediction model includes the following specific steps: An ecological governance effect prediction model is constructed, which includes a 3D dilated convolution module and a hierarchical temporal module, a multi-scale feature fusion layer, a 3D high-precision upsampling layer, and a prediction output layer. The 3D dilated convolution module includes: dilated convolution layer 1, dilated convolution layer 2, and a spatial attention module; Hollow convolution layer 1: Convolution kernel 3 3 3. Drillion ratio (1,2,2), number of output channels 128; dilated convolutional layer 2: convolutional kernel 3 3 3. Diffusion rate (1,2,2), number of output channels 128; Spatial attention module: Channel attention: global average pooling, global max pooling, shared multilayer perceptron (MLP), sigmoid activation to generate channel weights; Spatial attention: using channel average and max pooling to obtain 2-channel features, 7×7 convolution, sigmoid activation to generate spatial weights. The hierarchical temporal module includes: a short-term branch and a long-term branch. The short-term branch adopts a 3-layer convolutional long short-term memory network 3D-ConvLSTM, and the long-term branch adopts a 4-head attention layer and a feedforward network. 3D high-precision upsampling layer: transposed convolution: 3×3×3 kernel, stride (1,2,2), 128 channels, output 100×100×128; double transposed convolution: 3×3×3 kernel, stride 2, 64 channels; Prediction Output Layer: Ecological Governance Prediction Tensor , H represents the number of grid rows, W represents the number of grid columns, and K represents the ecological parameter channels. To predict the time step; Step S5: Collect real-time ecological governance data of the photovoltaic area to obtain the real-time ecological governance tensor; calculate the mean square error of ecological governance by calculating the parameter difference between the real-time ecological governance tensor and the predicted ecological governance tensor; calculate the channel similarity index of ecological governance by calculating the channel similarity between the real-time ecological governance tensor and the predicted ecological governance tensor. Step S6: By combining the mean square error of ecological governance and the similarity index of ecological governance channels, the ecological governance error index is calculated; based on the ecological governance error index, the parameters of the ecological governance effect prediction model are optimized, and finally the optimized ecological governance effect prediction model is obtained, thereby realizing the prediction of ecological governance effect.

2. The method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence according to claim 1, characterized in that: The process of filling in each point in the two-dimensional grid of the photovoltaic area using the inverse distance weighting method based on the soil data of the key area to calculate the soil data of the photovoltaic grid points includes the following specific steps: The soil data for each point in the two-dimensional grid of the photovoltaic area is calculated by filling in the grid using the inverse distance weighting method, based on the j-th feature value in the soil data of the key area: ； in, This represents the j-th feature value of the soil data at the photovoltaic grid point (h, w) in the two-dimensional grid of the photovoltaic area. Let (h, w) be the distance from the coordinates (h, w) to the soil data of the b-th key region, where b represents the index of the soil data of the b-th key region, and B represents the total number of soil data in the key regions. This represents the j-th feature value of the soil data in the b-th key area.

3. The method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence according to claim 2, characterized in that: The process of obtaining red band reflectance and near-infrared band reflectance by frequency band decomposition of the remote sensing image data includes the following steps: In UAV remote sensing image data, each pixel records the DN value for each band. The DN value is then converted into surface reflectance. ； in, For surface reflectance, The radiance is converted from the DN value using sensor gain and offset. The distance between the Earth and the Sun. Solar irradiance, The solar zenith angle; Based on surface reflectance data, the reflectance in the near-infrared band (700-1100nm) of UAV remote sensing imagery is extracted. Reflectivity in the red light band of the 600-700nm range .

4. The method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence according to claim 3, characterized in that: The calculation of the vegetation index at the photovoltaic grid points by combining the red light band reflectance and the near-infrared band reflectance includes the following steps: By combining the red light band reflectance and the near-infrared band reflectance, the vegetation index of the photovoltaic grid points is calculated: ； in, This represents the vegetation index at the photovoltaic grid point with coordinates (h,w). This represents the near-infrared reflectance at coordinates (h, w). This represents the reflectivity in the red light band at coordinates (h, w).

5. The method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence according to claim 4, characterized in that: The process of inputting the photovoltaic grid point synthesis tensor into the convolutional neural network model and outputting the ecological feature evolution tensor includes the following specific steps: The photovoltaic grid point synthesis tensor X is input into the convolutional neural network model, and the ecological feature evolution tensor is output. , Where H' and W' are the number of grid rows and columns output by the convolutional neural network model, T is the time step, and D is the ecological parameter channel output by the convolutional neural network model.

6. The method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence according to claim 5, characterized in that: The step of calculating the mean square error of ecological governance by calculating the parameter difference between the real-time ecological governance tensor and the predicted ecological governance tensor includes the following specific steps: The mean square error of ecological governance is calculated by determining the parameter difference between the real-time ecological governance tensor and the predicted ecological governance tensor. ； RMSE represents the mean square error of ecological governance. To predict the time step, t is the time step index, H is the number of grid rows, W is the number of grid columns, h is the index of the grid row number, and w is the index of the grid column number. This represents the ecological governance prediction tensor with h rows and w columns. This represents a real-time ecological governance tensor with h rows and w columns.

7. The method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence according to claim 6, characterized in that: The process of calculating the channel similarity index of ecological governance by measuring the channel difference between the real-time ecological governance tensor and the predicted ecological governance tensor includes the following specific steps: By calculating the channel similarity between the real-time ecological governance tensor and the predicted ecological governance tensor, the ecological governance channel similarity index is obtained: ； Among them, SSIM is the similarity index of ecological governance channels. To predict the time step, t is the time step index, K is the number of channels, and k is the channel index. Let be the mean of the predicted values ​​for the k-th channel at time step t. Let be the average of the measured values ​​of the k-th channel at time step t. Let Variance be the predicted value variance of the k-th channel at time step t. Let V be the variance of the measured values ​​of the k-th channel at time step t. Let be the covariance of the k-th channel at time step t. To prevent the mean from being divided by zero, the default value is... , The variance is divided by zero constant, which is the default value. .

8. The method for predicting the ecological governance effect of photovoltaic areas based on artificial intelligence according to claim 7, characterized in that: The ecological governance error index is calculated by combining the mean square error of ecological governance and the similarity index of ecological governance channels, including the following specific steps: By combining the mean square error of ecological governance and the similarity index of ecological governance channels, the ecological governance error index is calculated: ； Among them, EGDV is the ecological governance error index. To normalize the root mean square error, = MAXSE represents the maximum root mean square error, H is the number of grid rows, W is the number of grid columns, h is the index of the grid row, and w is the index of the grid column. The mean square error weighting coefficient is... The channel similarity index weighting coefficient, + =1.

Images